Spaces:
Sleeping
Sleeping
[agent:cc] org-card: chart is single-stream only - filter out concurrency>1 rows (no conc-aggregate mixing)
Browse files- hub_charts.py +5 -0
hub_charts.py
CHANGED
|
@@ -49,6 +49,11 @@ def _candidates(rows: list[dict], variant: str) -> list[dict]:
|
|
| 49 |
continue
|
| 50 |
if r.get("vram_used_gb") is None or r.get("tok_s_decode") is None:
|
| 51 |
continue
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
key = (r.get("model"), r.get("gpu_count"))
|
| 53 |
if key not in by or (r.get("tok_s_decode") or 0) > (by[key].get("tok_s_decode") or 0):
|
| 54 |
by[key] = r
|
|
|
|
| 49 |
continue
|
| 50 |
if r.get("vram_used_gb") is None or r.get("tok_s_decode") is None:
|
| 51 |
continue
|
| 52 |
+
# This chart is SINGLE-STREAM speed / power / eff. concurrency>1 rows are an
|
| 53 |
+
# aggregate-throughput measurement and must not be deduped against (or outrank)
|
| 54 |
+
# the single-stream point - else a model is plotted at its concurrent aggregate.
|
| 55 |
+
if (r.get("concurrency") or 1) != 1:
|
| 56 |
+
continue
|
| 57 |
key = (r.get("model"), r.get("gpu_count"))
|
| 58 |
if key not in by or (r.get("tok_s_decode") or 0) > (by[key].get("tok_s_decode") or 0):
|
| 59 |
by[key] = r
|